On the proof of efficacy of functional foods: design considerations

The ultimate objective of functional food science is to formulate hypotheses to be tested in human intervention studies that aim to show that the relevant intake of specified food components is associated with improvement in one or more target functions, either directly, or in terms of a valid marker of an improved state of health and well-being and/or a reduced risk of a disease. Functional food research encompasses several types of study designs, including observational studies and randomised clinical trials (RCTs). A sound scientific evidence from such studies is required to substantiate health claims on a functional food. A direct measurement of the effect a food on health is often not possible. Therefore, one key, but difficult, step in the development of functional foods is the identification and validation of relevant markers that can predict potential benefits or risks relating to certain health conditions. The second key step, once the surrogate has been identified and validated, is to frame the proof within a study design that is able to link the inference on the surrogate with the inference that would be made on the true endpoint, had it been observed. The aim of this work is to address this research question, by assuming that there is a known relationship between the surrogate and the true endpoint explored during the surrogacy assessment. In case of continuous ad normally distributed endpoints, design considerations in terms of sample size and power of a test on the true endpoint, based on its conditional variance under the established relation, are derived. We illustrate this approach through a motivating example from literature on surrogates for cardiovascular risk prevention. To address the issue of predictive accuracy, we integrate the example within a simulation study where an hypothetical shrinkage factor (with an assigned between-study distribution) is applied to correct for overestimation of the conditional expectation of the true endpoint. The power of a test on the mean of the true endpoint with shrinkage is compared to that without, under different data generating mechanisms. The joint consideration of the two sets of hypotheses, one on the surrogate and the other on the true endpoint, helps orienting the investigators and, hopefully, discourages RCTs on the surrogate that would allow to test only unrealistic effects on the true endpoint. As shown in the simulation study, over-optimistic prediction in the surrogate-true relation could result in a seriously underpowered test on the true endpoint.